17 posts
A scraping bridge asked the database for a column it never had. Nine days of runs, exit code zero, nothing collected. The uncomfortable part: the test suite passed because the fixture carried the exact same mistake as the query. Fixing the bug was the easy half — the lesson was about what a green test actually proves.
An LLM call ends in 200 OK with empty content and the pipeline moves on like nothing happened. Diagnosis: reasoning models burn the token budget thinking and nothing is left for the answer. The fixes: a token ceiling with real margin, and a parser that tolerates a data: [DONE] glued to the JSON.
The new stage in my transcription pipeline uses Gemini TEXT over 120s windows — no audio, cheap and fast. What keeps the model from making things up isn't a polite prompt: it's a numeric ±35% character invariant per window, with retry and flag when the math doesn't close.
After the bilingual jump to 0.1614 recall@1, there was a number left on the bench: how much was the cross-encoder contributing on every search? The week I measured that latency and recall are a bill you pay on every query — and the answer ended up written as a permanent flag in the config.
300 training steps, a loss stuck at 10.8125, and a model that insisted on coming out blank. The cause wasn't a low learning rate — it was exploded weights in the starting checkpoint. The story of how a stalled run became a full diagnosis.
A study about cron jobs: why they accumulate, fail silently, and how a cron that 'never sleeps' can take down an ecosystem. From Yurumi's consolidation to 289 orphan jobs — the lessons from a night of investigation.
Three embedding models, two search systems, one homegrown benchmark: the study that started with a question about semantic similarity and ended up revealing the tradeoffs between colibri, bge and openai — and how each one found its place in the ecosystem.
Capivara kept its memories in a local ChromaDB, with its own embedder and L1-L4 layers. Yurumi kept the ecosystem's memory in a Qdrant. Two separate memories meant a brain that never talked to the rest of the body. The unification on 08/28 migrated 182/182 memories without losing a single one: hash dedup, backup before deleting, and an --apply gate that separates the plan from the execution.
Every project's AGENTS.md was the official instruction set of the ecosystem, but it lived outside the Yurumi's memory — the ingest only ran when someone remembered to run it. The fix was a signature watcher (mtime+size) that re-ingests only what changed. Along the way: a first_run that ran a full ingest every day, an os.stat that hung when WSL wedged, and the lesson of never trusting a junction.
Yurumi stored its memory with a single embedding model trained in Portuguese. It worked — until the first question in English. The fix was not swapping the model; it was running two of them side by side and fusing the rankings with RRF. After a full LoCoMo run, recall went from 0.139 to 0.1614.
The simple question 'what does my agent actually remember?' turned into a full audit of Yurumi: 32 knowledge bases, 551 documents, a Qdrant collection with over 8 thousand vectors, and a LoCoMo baseline that handed back reality in percentage form.
I actually read the BitNet b1.58 paper and got stuck on a math problem that refused to add up: how do three values (-1, 0, +1) fit in 1.58 bits? A study week that became a didactic clone of ternary quantization — and wrecked my mental model of 'precision equals quality'.
After the wave field theory and Mamba post, the theory gained something it was missing: a didactic, portable, testable artifact. How to turn a study into code that runs — and why that changed how I learn.
A full inventory of the ecosystem's skills revealed 308 skills across 39 categories, 4 reuse gaps, and the padroes-reuso-ecossistema hub that organized everything into 7 domains. A study that turned into an audit of my own knowledge.
From the wave field theory paper to BitMamba-2 1B: how a study on wave propagation became the TatuEngine inference engine, and why the Ollama truncation loop (num_ctx 4096 vs 262K) became the Docs First rule that runs across the entire ecosystem.
Out of 78 posts, only 1 was about studies — the Ollama truncation loop. But it was in the studies that wave field theory, Mamba SSM and the BitMamba-2 1B were born. The story of what the blog's most neglected section actually produced.
Hermes kept entering a continuation loop 4 times before giving up. The cause: Ollama, by default, only allocates 4096 context tokens — regardless of the model supporting 262K. The hard way: num_ctx of the model ≠ num_ctx of the inference slot.